Re: [patch/rfc] perf on raspberry-pi without overflow interrupt

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Jan 13, 2014 at 11:55:17PM -0500, Vince Weaver wrote:
> On Fri, 10 Jan 2014, Peter Zijlstra wrote:
> 
> > On Thu, Jan 09, 2014 at 11:08:47PM -0500, Vince Weaver wrote:
> > > On Thu, 9 Jan 2014, Will Deacon wrote:
> > > 
> > > > I'd rather see it in the generic code if at all possible. Maybe we could add
> > > > a flags field to perf_pmu_register?
> > > 
> > > I can look into adding the check in generic code.
> > 
> > Adding something like this to the generic code would mean adding a
> > struct pmu capabilities field and visiting all existing PMU
> > implementations to properly fill this out.
> 
> I don't see an existing pmu capabilities struct... or do you mean
> coming up with one?

Yeah, adding one.

> Would it only hold an "overflow_interrupt_available" flag, or are
> there other generic capabilities it would be handy to know about?

Possible (other) flags could be:

  PMU_HAS_INT      -- would allow sampling events
  PMU_HAS_PRECISE  -- would allow any ::precise value
  PMU_HAS_FILTER   -- would allow all os/user/etc. flags

> > There's a number of hardware PMU implementations that do not have an
> > interrupt and would need to set this flag.
> 
> Well that can be added gradually, right?  Things wouldn't get any worse if 
> we add a generic check without auditing all code, things will just behave 
> the same as before for those architectures.

Right, doing a sweep once every so often is useful to find more patterns
though.

> There is some subtlety here though.  On ARM (or at least rasp-pi) the 
> overflow hardware is there, just no interrupt is hooked up.  So things 
> like counter overflow are handled as long as overflows aren't faster than 
> context switch time.  It's just sampled events aren't possible.
> 
> On architectures without overflow support at all (I've had such hardware; 
> some SPARC machines, the Playstation 3 hypervisor) then counter overflow 
> isn't possible without a periodic timer (sort of like what is done with 
> Intel uncore).  Is that something that should be in generic code too?

Maybe yeah, if there's enough replication of this it certainly makes
sense to lift it into generic code.